1. What does a "feedback loop" mean in the context of AI systems producing biased outcomes?
Right. A feedback loop is when the AI's current decisions shape the data it will learn from next — bias becomes self-reinforcing over time.
A feedback loop is about consequences feeding back into training data. The AI's biased decisions create a world that looks like the bias was correct, which the next version of the model then learns from.
2. The "specificity trap" means that in AI outputs, a statistic of "72.4%" is sometimes more suspicious than "about 70%". Why?
Correct. Decimal specificity is a learned credibility signal — AI generates it because it appears in authoritative sources. When there's no actual dataset, that specificity is a hallucination artifact.
There's no reliability threshold, AI can generate decimals, and "about 70%" isn't always from real research. The point is that decimal specificity signals credibility but doesn't guarantee accuracy.
3. Stanford's Foundation Model Transparency Index (2023) scored AI models on 100 dimensions. What was the significance of making the scoring rubric public?
Correct. A public rubric lets the audit be audited — readers can check specific scores, not just accept the overall rankings on faith.
Publishing the rubric made the audit itself auditable. Readers could check any specific score, which is exactly the standard of transparency the study was measuring AI companies against.
4. The lesson defines "contestability" as one of the three requirements for genuine AI accountability. Which of the following is an example of real contestability?
Contestability requires that affected individuals have a real, accessible path to challenge a decision — not a theoretical one. The Dutch SyRI ruling established that affected people have a legal right to understand the basis for automated decisions. That is contestability. An informational email or a general policy statement is not.
Contestability, as defined in the lesson, means affected individuals can actually challenge a decision through a process with real consequences. Emails, web pages, and terms of service don't provide that. The Dutch SyRI ruling — which established a legal right to understand and challenge automated decisions — is the model.
5. Steven Schwartz, the lawyer fined in 2023, went back to ChatGPT and asked it to confirm the fake cases were real. The AI confirmed them. What does this demonstrate?
Right. Double-down confirmation is a known failure mode. The model is responding to the conversational context, not rechecking its sources.
AI doesn't have intentions. It generates confident-sounding confirmations because that fits the conversational pattern — not because it found new evidence.
6. Why does an AI language model sometimes get simple arithmetic wrong, such as calculating the wrong percentage?
Right. The model predicts what a correct calculation result would look like, based on patterns — it doesn't run the numbers the way a calculator does.
The core issue is that text prediction is not computation. The model generates results that fit the pattern of correct answers — sometimes successfully, sometimes not.
7. A "finding" differs from a "verdict" in an audit because:
Correct. The distinction is critical: findings are data-grounded; verdicts go further into interpretation and judgment. Both have a place — they just need to be clearly labeled as different things.
Findings describe what data shows. Verdicts add interpretation — intent, moral weight, recommendations — that go beyond the data. The lesson uses the Stanford vs. journalism example to illustrate this distinction.
8. The Ofqual algorithm was reversed within 11 days. Which factor was most directly responsible for that speed?
Correct. Specific, numerical, publicly verifiable evidence moved faster than any complaint or petition could have. That's the leverage of a real audit.
The speed came from specific, numerical evidence drawn from publicly available data. The lesson describes it as an "informal but fast audit" — the methodology was what made it effective.
9. What is "proxy discrimination" in the context of AI systems?
Correct. Proxy discrimination uses seemingly neutral variables — school name, zip code, word choice — as indirect measures of characteristics the system is not supposed to use.
A proxy variable stands in for something else. Proxy discrimination is when a neutral-seeming variable effectively substitutes for a protected characteristic like race or gender.
10. A DOI (Digital Object Identifier) formatted like "doi:10.1016/j.jadohealth.2023.04.012" appears in an AI-generated citation. The DOI format looks correct. What should you do?
Correct. AI learns DOI formatting from millions of documents. It can generate correctly formatted DOIs that resolve to nothing. Only the actual lookup confirms existence.
Format correctness is not proof of existence. AI generates correct-format DOIs from pattern learning — the actual resolution test is the only reliable check.
11. The UK A-Level standardisation algorithm was reversed in August 2020 after nine days of public pressure. What element was most essential to making that reversal happen quickly?
The lesson is explicit: this is the model for successful AI accountability — technical literacy (researchers), communication (journalists), and informed affected parties (students) combining to create pressure that was specific enough to be politically unsustainable for the government to resist.
The lesson identifies this specifically. The reversal was fast because it combined technical expertise, journalistic communication, and informed affected students — not any single factor. That combination made the criticism specific and credible rather than just emotional.
12. What does RLHF — Reinforcement Learning from Human Feedback — do that makes it more robust than Tay's approach?
Correct. RLHF puts trained human raters between user interaction and model learning — instead of learning from all user input directly, the model learns from curated human evaluations.
RLHF adds structured human judgment as the learning signal. The model learns from what human raters approve of, not from direct user input — that is the key difference.
13. Which of the following best defines "training data bias"?
Correct. Training data bias is specifically about the AI inheriting and reproducing real-world inequalities embedded in its training data.
Training data bias is about what the data contains — reflections of real-world inequality — not about size, speed, or intentional error.
14. Joy Buolamwini's Gender Shades project tested which three commercial AI systems?
Correct. Buolamwini tested IBM, Microsoft, and Megvii — three major commercial facial recognition APIs available at the time.
The three systems were IBM Watson Visual Recognition, Microsoft Face API, and Megvii's Face++.
15. Robert Williams was arrested in 2020 because a facial recognition system returned a false match. What should have happened — but did not — before his arrest warrant was issued?
Correct. The failure was insufficient human verification — the AI's output was treated as conclusive rather than as a lead to be investigated.
The core failure was about human verification. The AI produced a candidate match; a human should have confirmed it before any arrest was made.
16. In August 2020, what did the UK's Ofqual algorithm do that caused a national controversy?
Correct. The algorithm used school-level historical performance to make individual-level predictions, systematically downgrading students at lower-performing state schools.
The algorithm downgraded students based on their school's historical performance — a group-level metric applied to individuals — and was reversed within 11 days after a fast statistical audit.
17. Microsoft's Tay chatbot was shut down in 2016 after sixteen hours. The key principle it illustrated about training data is:
Tay demonstrates that "learning from data" means learning from everything in the data — including coordinated bad-faith inputs. The system had no moral reasoning; it had pattern matching. That distinction is central to understanding what AI systems can and cannot do.
The key insight is more fundamental than platform choice. Tay illustrates that an AI learning from real human behavior inherits that behavior's worst aspects because it has no moral framework to evaluate what it's learning — only pattern recognition.
18. CNET's AI-written financial articles contained errors like calculating 3% as 3.3%. Why were these errors especially hard for readers to catch?
Correct. Subtle errors + authoritative presentation + reader trust in published content = a combination that allows errors to propagate undetected.
The combination of small numerical deviation, professional presentation, and normal reader trust in published financial journalism explains why the errors ran for two months before detection.
19. What is a confounding variable, and why does it matter in an AI audit?
Correct. Confounds are alternative explanations. Good audits name what they controlled for and acknowledge what they couldn't eliminate.
A confound is an outside factor — like school funding in the Ofqual case — that could explain findings without bias being the cause. Ruling out confounds is essential to credible audit conclusions.
20. A large language model trained primarily on English-language web content is used globally. What is the most significant risk that most users will never be aware of?
Representation in training data shapes what the system treats as normal, default, or true. A model trained predominantly on English content will treat English-speaking cultural frames as universal — silently, and with the same confidence it applies to everything else.
The technical issue (language support) is less important than the cultural framing issue. A model trained on predominantly English content will treat English-speaking cultural assumptions as the universal default — without any signal to users from other cultural contexts that this is happening.